Distributed simulator for stressing JobTracker and NameNode

Hadoop need a tool for stressing JobTracker and NameNode. Mumak introduced a simulated JobTracker, whose behavior doesn't exactly like that of the real JobTracker. Even more, mumak can't simulate a large cluster with quite a lot of jobs run on it. On the other hand, Gridmix v3 need hundreds of physical nodes to replay job stories.

You can think this tool a complementation of mumak and gridmix v3. We successfully used this tool to simulate a 12000 nodes cluster through 4 real machines.
I've talk to Hong Tang and Scott Chen offline, they suggested me contributing this tool to the hadoop community.